by Dariel Cruz Rodriguez, Rielley McNeill, and Derek Schneider | Website created with use of Claude Code
This project investigates whether significant differences exist in volatility of the sports prediction markets between men's and women's NCAA March Madness tournaments in 2026. We collect betting odds at regular intervals throughout games by scraping the Polymarket odds through their APIs to measure how dramatically odds swing as games progress on the day of play, in 5 min intervals. Our hypothesis predicts that women's tournament games will exhibit higher volatility due to reduced betting volume, but in fact it is actually men's games which ended up being more volatile. By analyzing odds fluctuations between both tournaments, we aim to quantify these volatility differences and explore their implications for sports betting markets. The findings are presented through an interactive website that allows users to explore the data across different games and time periods.
We collected betting odds data by web scraping the Polymarket APIs. First, we created a spreadsheet organizing the URLs for 125 Men's and Women's 2026 March Madness games (63 Men's games and 62 Women's games - not including the 4 games in the First Four opening round for each group). We then built code to extract betting odds data at 5-minute intervals for each game and generated individual CSV files to contain this information. These files were later merged into one large dataset containing all of the games. To perform our data cleaning process, we standardized column names across all the individual game datasets, ensured each game used the same general outcome betting market, and double checked that the merging of all game datasets into one combined dataset was executed properly. To measure volatility, we calculated metrics including probability range, maximum probability jumps, direction changes, trading volume, and the coefficient of variation (CV = probability standard deviation / mean probability x 100), our primary measure of normalized volatility. Finally, we designed an interactive website that visually resembles a March Madness bracket. Users can click on individual games to explore detailed statistics and volatility metrics for both Men's and Women's tournaments.
During the data collection part of our project, we ran into issues with the availability of betting odds data from the Polymarket API. We originally thought that because a significant amount of time has passed since the NCAA March Madness tournaments took place, the live in-game betting odds data would no longer be available through the API. However, we were still able to initially web scrape hour-by-hour betting odds data for the dates of the games. After further investigation into the API structure, we discovered that because March Madness markets remained open for less than 41 days, we could retrieve data at much finer 5-minute intervals. This substantially improved the granularity and quality of our collected data. Another challenge that came up involved differences in accessibility between Men's and Women's game pages. The URLs for Women's games were much more difficult to find, and the actual pages themselves were slightly different than those of the Men's games. Only graphs were present for Women's game pages while Men's game pages had actual odds data and a graph visible. By documenting these differences rather than ignoring them, we avoided making conclusions without acknowledging potential structural inequalities in how sports markets are presented. We also were able to address issues with missing outputs for certain March Madness games. Initially, some game URLs did not return any data when processed through our code. Instead of discarding large chunks of the dataset immediately, we carefully investigated potential causes, including API inconsistencies, broken market pages, and even checked to see if something simply went wrong with the way we entered the URL information into the spreadsheet organizer. Ultimately, we were able to recover data for nearly all the games, and excluded only one game due to incomplete data (UTSA vs UConn Women's)
Even though we were able to improve our dataset to capture 5-minute intervals, our original goal was to collect betting odds data in the form of continuous live-game updates (minute-by-minute). This could possibly reduce the precision of our comparisons between men's and women's games. As a result, some rapid market fluctuations still may not be fully captured in our analysis. Another limitation that emerged during our web scraping process was the limited accessibility of all bets placed. Because of legal and platform restrictions, our dataset only includes prediction market activity and bets from users based in the United States. So, our findings may not be an accurate reflection of global betting behavior. Finally, our analysis relied exclusively on the betting odds data from Polymarket. Our findings may not generalize to other sports betting or prediction market platforms.
If we had more time and resources to continue this project, we would want to further investigate whether or not prediction market sites like Polymarket have substantially different variability patterns when compared to traditional sportsbooks like DraftKings. Prediction markets operate through user-driven trading activity, while sportsbooks usually rely on odds that are set and adjusted by bookmakers. Comparing these systems could hopefully reveal whether volatility differences are influenced by market structure, trading volume, participant betting behavior, or platform design. This could also help us determine if differences in volatility patterns observed in Men's and Women's March Madness markets are unique to prediction markets, or if they are present across sports betting markets more broadly.